AI Compute Footprint — an honest estimator

What does one AI response actually cost in energy, water, and carbon? Tune the assumptions yourself and watch the range move — the honest answer is "it depends," and this tool shows you on what.

A public science-communication prototype from NSF NCAR / RAL, exploring the energy–water–climate nexus of computing.

⚠️ Read this first — these are illustrative estimates, not measurements

Nobody outside an AI operator's data centers knows the exact energy, water, or carbon cost of a specific query — the companies that run these systems rarely disclose exact per-query figures. Independent public estimates for a single response vary by one to two orders of magnitude depending on model size, hardware, prompt/response length, data center design, and the assumptions an analyst chooses.

This tool models the inference cost of using an already-trained model. Training a large model is a separate, very large, one-time energy/water/carbon cost, amortized over millions of later queries — it is not included in the numbers below.
Every number here is shown as a low–central–high range, never a single figure. Treat the central value as "a reasonable guess under stated assumptions," not a measurement.
Change the sliders below — that's the point. If the answer swings wildly, that's the honest lesson, not a bug.

Tune the assumptions

Every number on the right is computed live from these inputs.

Scenario presets

Scenario: Quick question

Model & response Model size / type

Response length

Session Messages in this session 1

Data center & grid assumptions Power Usage Effectiveness (PUE) 1.20 Overhead for cooling, power delivery, lighting on top of compute energy. 1.1–1.2 = efficient hyperscale facility; 1.5–1.6 ≈ global fleet average. On-site cooling water (WUE) 0.50 L/kWh Water evaporated on-site for cooling per kWh of IT energy. Best disclosed facilities: ~0.2; commonly cited averages: ~1.8–1.9. Off-site water (electricity generation) 1.50 L/kWh

Water consumed upstream to generate the electricity itself (evaporation at thermal/hydro power plants). Widely considered the most uncertain and most often omitted part of "AI water" figures. Grid carbon intensity 400 gCO₂e/kWh

Hydro/nuclear/wind-heavy grids (France, Norway, Quebec) run near the low end; coal-reliant grids (parts of India, Poland, some US regions) run near the high end.

This query

One response, given the assumptions on the left.

⚡ Energy

– – – – – Wh

low – central estimate, tune assumptions – high

💧 Water

– – – – – mL

low – central estimate, tune assumptions – high

on-site cooling ≈ – · off-site (generation) ≈ – (central values)

🌍 Carbon

– – – – – g CO₂e

low – central estimate, tune assumptions – high

Full session (1 message)

⚡ Energy

– – – – – Wh

low – central estimate, tune assumptions – high

💧 Water

– – – – – mL

low – central estimate, tune assumptions – high

on-site cooling ≈ – · off-site (generation) ≈ – (central values)

🌍 Carbon

– – – – – g CO₂e

low – central estimate, tune assumptions – high

What does that feel like? (approximate, illustrative equivalents from the central estimates)

This query, roughly...

This whole session, roughly...

A small picture

≈ 0 cups of water (session)

Compare the three preset scenarios

Session totals, central estimates only, computed with your current data-center/grid sliders on the left.

Scenario	Energy (Wh)	Water (mL)	Carbon (g CO₂e)

Methodology & sources — every default, its range, and where it comes from

This is the whole point of the tool: nothing here is hidden. If you disagree with a default, that's a feature — move the slider.

Parameter	Default used here	Plausible published range	Kind of source	Caveat
Energy per response — small / on-device model	0.02 Wh (central)	0.005–0.1 Wh	Order-of-magnitude inference from edge-hardware power draw; essentially no public per-query disclosures exist at this scale	Least-documented tier in this tool — treat as illustrative only
Energy per response — mid-size chat model	0.3 Wh (central)	0.1–1 Wh	Operator-disclosed figures (Google: ~0.24 Wh median Gemini text prompt; OpenAI/Sam Altman: ~0.34 Wh average ChatGPT query), as reported via IEA's "Energy and AI" analysis	These are GPU/compute-only figures; IEA notes the true facility total could be roughly double once cooling, networking and idle capacity are counted
Energy per response — large frontier model	3 Wh (central)	1–5 Wh	Independent academic/analyst estimates (de Vries, 2023, Joule; EPRI/BestBrokers analyses, 2024, ~2.9 Wh)	Independent estimates often run higher than recent operator disclosures — itself a demonstration of how assumption-dependent this figure is
Energy per response — frontier reasoning / agentic model	15 Wh (central)	5–50 Wh	No independent measurements are publicly available; inferred from reporting that reasoning/agentic models emit many more tokens per answer than standard chat models	Highest-uncertainty tier in this entire tool
Energy per image — image generation	3 Wh (central)	1–11 Wh per image	Luccioni et al., 2023, "Power Hungry Processing: Watts Driving the Cost of AI Deployment?" (measured across several diffusion-style image models)	Image/video generation is not directly comparable to per-token text costs — shown separately by design
Response-length multiplier	short ×0.4 / typical ×1 / long ×2 / extensive ×4	—	Simplified linear proxy for output-token count	Real inference cost also depends on input/context length, batching, and hardware — not modeled here
PUE (data center overhead)	1.2	1.1–1.6	Uptime Institute global data center survey (fleet-wide averages have hovered near 1.5–1.6 in recent years)	Leading hyperscale AI facilities self-report nearer 1.1; older/smaller facilities run higher
On-site cooling water (WUE)	0.5 L/kWh	0.1–2.0 L/kWh	Operator sustainability reports (e.g., AWS ~0.19 L/kWh fleet average); other commonly cited industry figures sit near 1.8–1.9 L/kWh	Extremely climate- and cooling-design-dependent (evaporative vs. closed-loop vs. air cooling)
Off-site water (electricity generation)	1.5 L/kWh	0–4 L/kWh	Power-sector water-intensity literature (thermoelectric water-consumption factors, in the tradition of NREL/Macknick-style analyses)	Called out by researchers as the most uncertain, and most often entirely omitted, part of public "AI water footprint" claims
Grid carbon intensity	400 gCO₂e/kWh (US-average preset)	~50 (very low-carbon) to ~450 (global average) to ~700–800 (coal-heavy)	IEA Electricity 2025; national/regional grid emission-factor datasets	These are annual averages — real-time "marginal" emissions at the moment of use can differ substantially
Extra uncertainty bands applied to your slider choices	on-site ×0.5/×1.5, off-site ×0.3/×3, carbon ×0.75/×1.3 around whatever you set	—	Added because even a "known" input like PUE, WUE, or grid intensity is itself a disclosed estimate, not a physical constant	Low-with-low and high-with-high are combined across factors — a simple, transparent bounding method, not a formal statistical or Monte Carlo propagation
Equivalents (LED bulb, phone charge, tea, teaspoon, sip, bottle, km driven, gas stove)	LED bulb 9 W; phone charge 12 Wh; tea 25 Wh/cup; teaspoon 5 mL; sip 15 mL; bottle 500 mL; car ~170 gCO₂e/km; gas stove burner ~0.14 gCO₂e/s	—	Basic physics (specific heat of water, for "energy to boil a cup of tea") plus commonly cited reference values (typical LED wattage, smartphone battery capacity, average passenger-car emissions factor, gas combustion emission factors)	Meant as intuitive anchors to make numbers tangible, not precise conversions

How the ranges are combined: Energy per response (low/central/high) is multiplied by the response-length multiplier and by PUE to get facility energy, then by the number of messages for the session total. Water = facility energy (Wh) × on-site WUE (low/central/high) for on-site cooling, plus facility energy × off-site water intensity (low/central/high) for embodied generation water; the two are added. Carbon = facility energy (kWh) × grid carbon intensity (low/central/high). At every step, "low" combines with "low" and "high" with "high" — an intentionally simple, inspectable method chosen so the mechanism is auditable, not a black box.

Sources & further reading (kinds of sources these figures represent)

International Energy Agency — Energy and AI and Electricity 2025 reports (operator-reported figures, grid emissions data)
Epoch AI — "How much energy does ChatGPT use?" gradient-updates analysis (2025)
Public sustainability disclosures from Google, OpenAI, Microsoft, and Amazon Web Services (per-query energy/water figures, fleet-wide PUE/WUE)
de Vries, A. (2023). "The growing energy footprint of artificial intelligence." Joule — independent academic estimate
Li, P. et al. "Making AI Less Thirsty" (arXiv:2304.03271) — UC Riverside / UT Arlington academic water-footprint analysis
Luccioni, A. et al. (2023). "Power Hungry Processing: Watts Driving the Cost of AI Deployment?" — measured image-generation energy across models
Uptime Institute — annual Global Data Center Survey (PUE benchmarking)
NREL / Macknick et al.–style water-intensity-of-electricity-generation literature (power-sector water consumption factors)
US EPA / European Environment Agency — vehicle and combustion emission factors (used only for the driving/gas-stove equivalents)

All of the above figures are contested, method-dependent, and change as hardware and grids change. Treat this tool as a way to reason about orders of magnitude and trade-offs — not as an authoritative measurement of any specific AI product.